Feature: AI Stylistic Tell Detection — Tricolon, Stacked Sentences, and AI Vocabulary Fingerprints

[craigtrim]

Inspiration

From a widely-shared LinkedIn post by Cassidy Johnston:

"It's not the emdash, it's the word 'quiet' in some form in your post along with things in threes ('It's not this, it's that, or that.' 'Profound short statement. Other profound short statement. Mic drop.')

And sentences
Structured
Like this

That tell me that you're using AI to write your posts."

This is a practitioner-level observation that maps directly to measurable stylometric features. These are the tells that signal AI-generated text — not because they're impossible for humans to write, but because they appear with statistically anomalous frequency and co-occurrence in LLM output.

Identified AI Stylistic Tells

1. Tricolon / Rule of Three

LLMs have a strong bias toward three-part structures:

• "It's not this, it's that, or that."
• "First thing. Second thing. Third thing."

Detectable as: sentence-level syntactic tricolon; comma-separated list-of-three patterns within a sentence.

2. Stacked Single-Sentence Paragraphs

And sentences
Structured
Like this

Each "sentence" is a single word or short phrase on its own line, used for dramatic effect. LLMs overuse this pattern as a rhetorical device.

Detectable as: paragraph length distribution; ratio of 1-word and 2-word paragraphs; stacked ultra-short-sentence sequences.

3. AI Vocabulary Fingerprints

Certain words appear with statistically elevated frequency in AI-generated text:

• quiet / quietly
• delve, tapestry, nuanced, robust, pivotal, transformative
• it's not X, it's Y framing

Detectable as: word-level frequency anomaly against human baseline corpora; presence of known AI-preferred lexical items.

4. Mic Drop Sentence Structure

Pattern: [Profound claim]. [Restatement]. [Short punchy close].

Detectable as: sentence length sequence — long, medium, very short — appearing in final paragraph position.

Relationship to Existing Work

This directly extends:

• Feature: Add AI-Generated Text Detection Module #65 (AI-Generated Text Detection) — these are concrete, named features to detect
• Feature: Tonality Derivation and Generative AI Style Transfer Validation #68 (Tonality Derivation) — these tells define what to avoid when conformance-scoring LLM output against a human author

Proposed Implementation

• Tricolon detector: identify comma-list-of-three and three-sentence sequences
• Stacked short paragraph detector: sequences of 1–3 word paragraphs
• AI vocabulary frequency scorer: flag elevated use of known LLM-preferred words against a baseline
• Mic drop sentence shape detector: length-descending terminal sentence sequences
• Aggregate "AI tell score" combining the above signals
• CLI output and integration into mega-meta pipeline

Why This Matters

These features are not just useful for AI detection — they're useful for style coaching. A human writer who wants to sound less AI-generated can use these scores as a mirror. Conversely, when validating LLM output against a human author's tonality (#68), high AI-tell scores are a direct conformance penalty.

[craigtrim]

Detection Strategies

Pattern 1: Tricolon / Rule of Three

The signal is not presence but density — humans use tricolon deliberately and sparingly; LLMs use it as a default organizational skeleton.

Surface form (comma-list-of-three):

X, Y, and Z    (Oxford tricolon)
X, Y, or Z     (alternative tricolon)

Detectable with regex against sentence tokens.

Clausal form (parallel clause tricolon):
Sentences containing exactly two commas that divide the sentence into three roughly equal-length segments is a pragmatic proxy without needing a full syntactic parser.

Metric: tricolon_density = tricolon instances per 1000 words vs. a human baseline corpus. The anomaly is frequency, not existence.

---

Pattern 2: Mic Drop — Paragraph-Terminal Deceleration

The shape is a rhythmic ritardando — the text slows to a staccato stop at paragraph close:

[longer setup sentence]  [medium sentence]  [very short sentence]

Measurement approach:

1. Compute sentence word-counts within each paragraph
2. For the final 2–3 sentences, compare length against the paragraph mean
3. A mic drop signature = terminal sentences significantly shorter than paragraph average, with a monotonically decreasing or steeply below-average trend

Metric: terminal_brevity_ratio = last_sentence_length / paragraph_mean_sentence_length

A ratio below ~0.4 at paragraph-end, appearing with high frequency across a document, is a strong signal.

---

Pattern 3: Stacked Single-Sentence Paragraphs

And sentences
Structured
Like this

Metric:

• single_sentence_paragraph_ratio — proportion of paragraphs containing only one sentence
• short_paragraph_run_length — longest consecutive run of single-sentence paragraphs

---

Macro-Rhythmic Framing

These patterns are prosodic, not merely syntactic — they operate at the paragraph level and above, which is largely unexplored in computational stylistics:

Pattern	Musical analogue	Rhythmic effect
Tricolon	Ternary meter	Rising expectation, release on third beat
Mic drop	Ritardando + fermata	Deceleration into dramatic pause
Stacked lines	Staccato	Hammered, separated beats

---

The Compound Signal

The strongest AI tell is co-occurrence. A single paragraph may contain:

• A tricolon setup sentence
• Followed by a mic drop close

Measured as a joint probability across a document, this compound pattern is a far stronger signal than either feature alone. A human using one occasionally is stylistically unremarkable. Both together, repeatedly, is a fingerprint.

This should factor into the conformance penalty in #68 — high compound AI-tell density is a direct deduction against style fidelity to a human author.

[craigtrim]

Detection Implementation Map

What We Already Have

Three existing modules directly contribute to these detections without modification:

syntactic/sentence_stats.py

• Produces sentence_lengths list (in metadata), mean, std, min, max per document
• The raw material for mic drop detection — we just need to apply it at paragraph granularity rather than document level

prosody/sentence_syllable_patterns.py

• Per-sentence syllable counts and complexity_uniformity_score
• Already flagged as an AI detection signal (uniformity > 0.8)
• Contributes to stacked short paragraph detection via sentence-level syllable shape

prosody/syllable_pattern_repetition.py

• starting_pattern_repetition_rate and ending_pattern_repetition_rate — directly measure formulaic sentence openings/closings
• repetition_ratio — proportion of patterns that recur (higher = more formulaic/AI-like)
• pattern_entropy — Shannon entropy of pattern distribution (lower = more template-driven)

syntactic/sentence_types.py

• Simple/compound/complex classification
• High simple sentence ratio + short sentence length = strong stacked-sentence signal

---

How to Combine Existing Features (No New Code Required)

Mic Drop Score (paragraph-terminal deceleration):

sentence_stats.metadata["sentence_lengths"] 
  → segment by paragraph boundaries
  → terminal_brevity_ratio = last_sentence_length / paragraph_mean
  → score = mean(terminal_brevity_ratio) across paragraphs

Stacked Sentence Score (hammered short paragraphs):

sentence_types.simple_ratio 
  × (1 - sentence_stats.sentence_length_std / sentence_stats.mean_sentence_length)
  → high simple ratio + low variance = stacked pattern

Formulaic Rhythm Score (opening/closing repetition):

syllable_pattern_repetition.starting_pattern_repetition_rate
  + syllable_pattern_repetition.ending_pattern_repetition_rate
  + (1.0 - syllable_pattern_repetition.pattern_entropy / max_expected_entropy)
  → normalized compound score

Compound AI-Tell Score (aggregate):

sentence_syllable_patterns.complexity_uniformity_score   # syllabic uniformity
  + syllable_pattern_repetition.repetition_ratio          # formulaic patterns  
  + stacked_sentence_score                                 # paragraph structure
  + mic_drop_score                                         # terminal deceleration
  → weighted average → AI-tell confidence

---

What Still Needs to Be Built

Feature	Gap	Notes
Tricolon detector	No current module	List-of-three regex + equal-segment comma split as proxy
Paragraph-level segmentation	All analysis is sentence-scoped	Need a paragraph tokenizer to enable terminal brevity per paragraph
AI vocabulary wordlist	No LLM-preferred word frequency list	Words like "quiet", "delve", "tapestry", "nuanced" — needs a curated baseline
Co-occurrence scorer	No module joins multiple AI-tell signals	Aggregation layer that weights and combines existing + new signals

The paragraph segmentation gap is the most foundational — mic drop detection and stacked-paragraph detection both require it. That's the logical first build target.

[craigtrim]

Gap Issues Filed

The four implementation gaps identified above have been broken out into individual issues:

#	Title	Priority
#72	Paragraph-Level Segmentation	Highest — unblocks mic drop + stacked paragraph detection
#73	Tricolon Detector	New feature, no existing foundation
#74	AI Vocabulary Wordlist and Frequency Scorer	New feature, requires baseline corpus
#75	AI-Tell Co-occurrence Scorer	Aggregation layer — depends on all others

Recommended build order: #72 → #73 → #74 → #75

[craigtrim]

Paragraph Segmentation Dependency Is Resolved

Both blocked detectors (mic drop, stacked short paragraphs) depend on paragraph boundaries. That dependency is already satisfied by fast-sentence-segment.

segment() returns list[list[str]] — paragraphs as outer lists, sentences as inner. The current split_sentences() in _utils.py calls segment() and flattens the result, discarding paragraph structure. Exposing a split_paragraphs() wrapper (no new dependency, no new logic — see #71) unblocks both detectors immediately.

Detector Status After #71 Is Closed

Detector	Status	Dependency
Mic drop (terminal_brevity_ratio)	Unblocked	split_paragraphs() via #71
Stacked short paragraphs (short_paragraph_run_length)	Unblocked	split_paragraphs() via #71
Tricolon	Unblocked (see #70, #73)	spaCy already present; #80 for POS head-word check
AI vocabulary scorer	Unblocked (see #74)	No paragraph dependency

Blocking Order

1. Feature: Paragraph-Level Segmentation #71 — split_paragraphs() added to _utils.py (closes immediately, zero new work)
2. Feature: Tricolon Detector — Rule of Three Pattern Recognition #70 / Feature: Tricolon Detector #73 — tricolon detector (unblocked once Feature: Paragraph-Level Segmentation #71 is done; Feature: Head-Word POS Extractor for Phrase Strings #80 for head-word POS)
3. This issue (Feature: AI Stylistic Tell Detection — Tricolon, Stacked Sentences, and AI Vocabulary Fingerprints #69) — mic drop + stacked paragraph detectors (unblocked by Feature: Paragraph-Level Segmentation #71)
4. Feature: AI-Tell Co-occurrence Scorer #75 — co-occurrence scorer (aggregates all of the above)

Related

• Feature: Paragraph-Level Segmentation #71 — immediate prerequisite (can close now)
• Feature: Paragraph-Level Segmentation #72 — duplicate of Feature: Paragraph-Level Segmentation #71
• Feature: Tricolon Detector — Rule of Three Pattern Recognition #70 — tricolon detector (rule-of-three)
• Feature: Tricolon Detector #73 — tricolon detector (companion)
• Feature: AI-Tell Co-occurrence Scorer #75 — co-occurrence scorer (downstream aggregator)
• Feature: Head-Word POS Extractor for Phrase Strings #80 — head-word POS extractor (unblocks tricolon syntactic validation)
• craigtrim/fast-sentence-segment — source library providing native paragraph output
• NewlinesToPeriods silently corrupts text mid-word when hard-wrapped line breaks fall inside a word boundary fast-sentence-segment#42 — NewlinesToPeriods mid-word corruption on hard-wrapped plain text (affects paragraph boundary quality)
• Hard-wrapped ALL-CAPS block quote split mid-prepositional-phrase by NewlinesToPeriods fast-sentence-segment#43 — NewlinesToPeriods adjective/noun split (same class of issue)
• NewlinesToPeriods splits adjective from its noun when line break falls mid-phrase fast-sentence-segment#44 — NewlinesToPeriods adjective/noun split (same class of issue)